fork note:

Original repo is: https://github.com/jaesung-cs/vkgs Current top of branch uses 2 queus: one for dealing with .ply file transfer and one for drawing splats. My laptop has the only queue so I reverted commits and returned to one-queue solution.

pygs

Gaussian Splatting

Requirements

• submodules

$ git submodule update --init --recursive

,,,

• conda

$ conda create -n pygs python=3.10
$ conda activate pygs
$ conda install conda-forge::cmake
$ conda install nvidia/label/cuda-12.2.2::cuda-toolkit  # or any other version

Build

$ cmake . -B build
$ cmake --build build --config Release -j

Run

$ ./build/pygs_base

Notes

• Directly updating to vulkan-cuda mapped mempry in kernel is slower than running cuda kernel and then memcpy (3.2ms vs. 1ms for 1600x900 rgba32 image)